H3 video generation issue: phantom audio reads out prompt text
Vijayi · reddit · 2026-08-26
A user reported a glitch in generated videos (likely H3 model) where random voiceovers occur, reading phrases directly from the prompt or speaking unintelligible gibberish. This happens even when no speech is mentioned in the prompt. The user observed that camera direction instructions appended to the prompt are occasionally spoken aloud in the clip. Current settings: 0.4 guidance, 8 steps, baked-in LoRA checkpoint, 10-12s duration.
More from Multimodal
- The math: 5 rerolls at 480p plus a 4K upscale costs $5 vs $9 at 720p — HeyAmit_ · 2026-10-15
- Dreamina workflow: generate at 480p, upscale to 4K, cut video costs 44% — HeyAmit_ · 2026-10-15
- One-prompt GPT-6 Astra writes two Palestrina-style motets in LilyPond — deanwball · 2026-10-15
- DiffSynth-Studio Adds OPD Workflow: Few-Step Diffusion Training in ~2 GPU-Hours — bdsqlsz · 2026-10-15
- AI film and TV hit the market: Prime Video picks up AI spin-off, pilots screen at Mipcom — lmoroney · 2026-10-15
- Ramayana film taps DNEG's massive VFX and AI-assisted production for global release — CurieuxExplorer · 2026-10-15